Efficient Learning of Pseudo-Boolean Functions from Limited Training Data